Methodology for stiffness prediction in structural timber using cross-validation RMSE analysis

Non-destructive testing (NDT) methods are becoming increasingly popular for estimating the mechanical properties of structural timber without altering the samples. These methods yield parameters that can be integrated into multivariate models, improving predictive accuracy compared to univariate models. Unfortunately, overfitting is often neglected, resulting in models that fit the experimental dataset perfectly but perform poorly on new data. The aim of the present paper is to analyse model prediction for timber stiffness through cross-validated root-mean-square error (RMSE), oriented to support the results obtained using several NDT methodologies avoiding overfitting. The research involved 491 samples of Pinus pinaster Aiton structural sawn timber. The results pointed out that dynamic moduli of elasticity had the lowest prediction error for stiffness, regardless of the device employed. Furthermore, no statistically significant differences were identified according to the sensor arrangements on the devices based on time-of-flight. Artificial Neural Networks and Multiple Linear Regression were effective multivariate algorithms for stiffness prediction, with no statistically significant differences in their prediction errors. The device which combined the lowest prediction error and number of variables was the Portable Lumber Grader which uses dynamic modulus of elasticity and knottiness. In contrast, the device with the lowest predictive capacity was machine strength grading. • RMSE can be a better indicator than R2 because it explains the model accuracy. • Complex nonlinear models could not improve the prediction of the linear models. • PLG showed a lower error than Microsecond Timer, Sylvatest Duo and Cook Bolinder. • In ultrasounds devices direct and indirect measurements showed the same accuracy. • Cross-validation was a good strategy to avoid overfitting in multivariate models.